A knowledge graph-based intelligent recommendation method, system, and storage medium for tumor reagents.

Abstract

This application relates to the field of data processing technology and discloses a method, system, and storage medium for intelligent recommendation of tumor reagents based on knowledge graphs. The method includes: constructing a temporal dynamic knowledge graph by adding timestamp attributes to relevant nodes; updating the node memory state when an event occurs, generating a time embedding vector; acquiring patient data to identify TNM staging and outputting a state vector; calculating time-weighted semantic similarity to screen candidate reagents; optimizing the path based on TNM staging constraints, and generating recommendation results. This application solves the problems of static knowledge representation's inability to adapt to temporal changes and lack of personalized path optimization by constructing a temporal dynamic knowledge graph and using a Q-learning optimization algorithm based on TNM staging constraints, thereby improving the timeliness and personalized accuracy of tumor reagent recommendations.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method, system and storage medium for intelligent recommendation of tumor reagents based on knowledge graphs. Background Technology

[0002] In the field of tumor diagnosis and treatment, knowledge graph-based intelligent recommendation methods for tumor reagents have been gradually applied in clinical practice. Existing technologies construct static knowledge graphs containing medical entities such as tumor reagents, disease types, and genetic information. Using semantic similarity calculations and path search algorithms, they match and analyze the clinical features input by patients to recommend suitable tumor testing reagents. These methods, to some extent, address the inefficiencies and inaccuracies of traditional manual reagent selection, providing auxiliary decision support for clinicians.

[0003] However, existing knowledge graph-based tumor reagent recommendation methods have significant shortcomings. First, existing methods rely on static knowledge graphs and cannot handle temporal characteristics such as the decay of reagent detection accuracy over time and the dynamic fluctuations in biomarker expression levels, resulting in poor timeliness of recommendation results. Second, existing technologies lack in-depth utilization of patients' TNM staging information and fail to adjust recommendation strategies according to tumor severity. In addition, traditional methods lack intelligent path optimization mechanisms in the reagent selection process, making it impossible to achieve optimal decisions under complex multi-dimensional evaluation criteria.

[0004] The core problem with existing technologies lies in the contradiction between static knowledge representation and the dynamic needs of medical practice. The time-dependent decay of reagent performance requires knowledge graphs to be dynamically updated, while the differentiated characteristics of tumor staging necessitate personalized path optimization capabilities in recommendation algorithms. Therefore, how to construct a knowledge graph capable of capturing time-dependent dynamic changes and combining it with TNM staging constraints to achieve intelligent path selection has become a critical technical problem that urgently needs to be solved. Summary of the Invention

[0005] This application provides a knowledge graph-based intelligent recommendation method, system, and storage medium for tumor reagents. By constructing a time-series dynamic knowledge graph and employing a Q-learning optimization algorithm based on TNM staging constraints, it addresses the problems of static knowledge representation's inability to adapt to temporal changes and the lack of personalized path optimization in existing technologies. This invention improves the timeliness and personalized accuracy of tumor reagent recommendations.

[0006] Firstly, this application provides a knowledge graph-based intelligent recommendation method for tumor reagents, the knowledge graph-based intelligent recommendation method for tumor reagents comprising:

[0007] Step S101: Add timestamp attributes to the tumor reagent node, biomarker node, and gene node, establish the correlation between the reagent detection accuracy attenuation coefficient and the biomarker expression fluctuation rate, and construct a time-series dynamic knowledge graph.

[0008] Step S102: When a reagent detection event occurs, update the memory state of the nodes in the time-series dynamic knowledge graph and generate a node time embedding vector containing dynamic features and time decay features.

[0009] Step S103: Obtain the patient's clinical characteristics, gene detection results and protein expression data, identify the patient's TNM staging information and output the patient state vector containing the scores of each staging;

[0010] Step S104: Calculate the time-weighted semantic similarity between the node time embedding vector and the patient state vector, and filter reagent nodes with a correlation degree higher than the threshold to obtain a candidate reagent set;

[0011] Step S105: Optimize the selection path of the candidate reagent set based on TNM staging constraints, and generate intelligent recommendation results including reagent identification, comprehensive score and recommendation basis.

[0012] Optionally, step S101 includes:

[0013] We collected literature, reagent product instructions, and clinical databases in the field of oncology. We then cleaned and standardized the raw data to obtain a standardized basic dataset of oncology reagents.

[0014] Based on the standardized tumor reagent dataset, entity recognition and relation extraction are performed to obtain an entity set containing tumor reagent entities, biomarker entities, gene entities, and protein entities, as well as a set of relationships between entities.

[0015] Add a timestamp attribute and a decay coefficient attribute to each entity node in the entity set, and add a time weight attribute and an intensity change rate attribute to each relationship edge in the entity relationship set to obtain time-annotated graph elements;

[0016] The time-annotated spectral elements are stored in a graph database, and a decay function of reagent detection accuracy with time and a fluctuation function of biomarker expression level with time are established to obtain the time-series dynamic knowledge graph.

[0017] Optionally, step S102 includes:

[0018] The occurrence of reagent testing events in the time-series dynamic knowledge graph is monitored, and the event type, occurrence time, and performance data of the reagent testing events are identified and processed to obtain reagent testing event data;

[0019] The current detection accuracy of the relevant reagent nodes is calculated based on the reagent detection event data to reduce the detection accuracy. The updated detection accuracy value is calculated based on the time interval and the preset reduction rate to obtain the reagent node performance update data.

[0020] The performance update data of the reagent node is input into the memory state update module for node memory state refresh processing. At the same time, the expression level of the biomarker node is processed by fluctuation coefficient calculation to obtain the node update state containing performance decay information and expression fluctuation information.

[0021] Based on the node update state, feature fusion processing is performed to combine and encode the node's inherent attributes, temporal change features, and time decay features to obtain the node's time embedding vector.

[0022] Optionally, step S103 includes:

[0023] Collect basic patient information, medical history records, and symptom descriptions, and perform standardized coding processing on the patient's clinical characteristics to obtain a clinical feature vector;

[0024] Mutation information and expression profile data are extracted from the gene detection results, and the gene mutation types and expression levels are quantitatively analyzed to obtain gene feature vectors.

[0025] The protein concentration and immunohistochemical results in the protein expression data are obtained, and the protein expression level is normalized to obtain the protein feature vector.

[0026] The clinical feature vector, gene feature vector, and protein feature vector are input into the TNM staging classifier for staging identification, and tumor size score, lymph node involvement score, and distant metastasis score are calculated to obtain the patient status vector.

[0027] Optionally, step S104 includes:

[0028] Extract the reagent feature component and time decay component from the node time embedding vector, and perform vector alignment processing on the clinical component and gene component in the patient state vector to obtain feature vector pairs;

[0029] Based on the feature vector, the cosine similarity value is calculated. The time weight coefficient is calculated according to the timestamp attribute of the reagent node. The cosine similarity value is then adjusted by time weighting to obtain a time-weighted semantic similarity score.

[0030] A threshold for the degree of association is set to a preset value. The time-weighted semantic similarity score is compared with the threshold, and reagent nodes with scores higher than the degree of association are filtered to obtain a set of filtered reagent nodes.

[0031] The set of selected reagent nodes is sorted in descending order according to the time-weighted semantic similarity score, and the top N reagent nodes are selected to obtain the candidate reagent set.

[0032] Optionally, step S105 includes:

[0033] Define the state space to include the candidate reagent set and the patient state vector, set the action space to include reagent selection action and path switching action, establish TNM staging constraints, and obtain the reinforcement learning environment configuration;

[0034] The candidate reagent set is input into the Q-learning algorithm for state value function initialization. The reward function weight parameters are set according to the TNM stage score in the patient state vector to obtain the Q-value matrix.

[0035] Based on the Q-value matrix, an action selection strategy is executed, and a weighted sum of reagent detection accuracy reward, cost-effectiveness reward, and time-effect reward is calculated. The Q-value matrix is ​​then iteratively updated to obtain an optimized path sequence.

[0036] The optimal reagent selection scheme is extracted based on the optimized path sequence, and a comprehensive score is calculated for reagent detection accuracy, stability index, and cost-effectiveness ratio. The intelligent recommendation result, which includes reagent identification, the comprehensive score, and selection criteria, is then generated.

[0037] Optionally, the step of inputting the candidate reagent set into the Q-learning algorithm for state value function initialization processing, setting reward function weight parameters according to the TNM stage score in the patient state vector, and obtaining the Q-value matrix includes:

[0038] Each reagent node identifier in the candidate reagent set is mapped to a state number, and the tumor size score, lymph node involvement score, and distant metastasis score in the patient state vector are normalized to obtain a state space coding table.

[0039] The Q-value function matrix is ​​initialized based on the state space encoding table. The Q-value of each state-action pair in the matrix is ​​randomly initialized. The learning rate parameter and discount factor parameter are set to obtain the initial Q-value matrix.

[0040] Extract the TNM stage score values ​​from the patient state vector, calculate the detection accuracy weight, cost sensitivity weight, and time urgency weight according to the tumor severity, and perform normalization constraint processing on the weight parameters to obtain the set of reward function weight parameters;

[0041] The reward function weight parameter set is embedded into the update formula of the initial Q-value matrix to construct the state transition probability function and the instantaneous reward function of the TNM phased constraint, thereby obtaining the Q-value matrix.

[0042] Secondly, this application provides a knowledge graph-based intelligent recommendation system for tumor reagents, the knowledge graph-based intelligent recommendation system for tumor reagents comprising:

[0043] A module was established to add timestamp attributes to tumor reagent nodes, biomarker nodes, and gene nodes, establish the correlation between reagent detection accuracy attenuation coefficient and biomarker expression volatility, and construct a time-series dynamic knowledge graph.

[0044] The generation module is used to update the memory state of nodes in the time-series dynamic knowledge graph when a reagent detection event occurs, and generate node time embedding vectors containing dynamic features and time decay features.

[0045] The identification module is used to acquire patients’ clinical characteristics, gene testing results and protein expression data, and after identifying the patient’s TNM staging information, it outputs a patient state vector containing the scores of each staging.

[0046] The filtering module is used to calculate the time-weighted semantic similarity between the node time embedding vector and the patient state vector, and filter reagent nodes with a correlation degree higher than a threshold to obtain a candidate reagent set.

[0047] The optimization module is used to optimize the selection path of the candidate reagent set based on TNM staging constraints, and generate intelligent recommendation results including reagent identification, comprehensive score and recommendation basis.

[0048] Thirdly, a knowledge graph-based intelligent recommendation device for tumor reagents is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the knowledge graph-based intelligent recommendation device for tumor reagents to execute the aforementioned knowledge graph-based intelligent recommendation method for tumor reagents.

[0049] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the aforementioned knowledge graph-based intelligent recommendation method for tumor reagents.

[0050] The technical solution provided in this application utilizes a knowledge graph-based intelligent recommendation method for tumor reagents. This method adds timestamp attributes to tumor reagent nodes, biomarker nodes, and gene nodes, and establishes a correlation between the reagent detection accuracy decay coefficient and biomarker expression volatility, constructing a temporal dynamic knowledge graph. This fundamentally solves the technical deficiency of existing static knowledge graphs, which cannot reflect the temporal changes in medical information. When a reagent detection event occurs, the memory state of nodes in the temporal dynamic knowledge graph is updated, generating a node temporal embedding vector containing dynamic and temporal decay features. This allows the recommendation system to capture the decay trend of reagent performance and the fluctuation pattern of biomarker expression in real time, ensuring the timeliness and accuracy of the recommendation results. By acquiring patient clinical characteristics, gene detection results, and protein expression data, and identifying the patient's TNM staging information, a patient state vector containing scores for each stage is output, achieving a precise quantitative description of the patient's tumor severity and laying a solid foundation for subsequent personalized recommendations. The technical solution of calculating the time-weighted semantic similarity between the node temporal embedding vector and the patient state vector, and filtering reagent nodes with a correlation degree higher than a threshold to obtain a candidate reagent set, effectively integrates temporal information and semantic matching, and significantly improves the accuracy of candidate reagent screening.

[0051] The core technical feature of optimizing the selection path of candidate reagent sets based on TNM staging constraints and generating intelligent recommendation results including reagent identification, comprehensive scores, and recommendation criteria fully demonstrates the specific application advantages of reinforcement learning algorithms in the field of tumor reagent recommendation. By embedding TNM staging constraints into the state transition and reward mechanism of the Q-learning algorithm, the path optimization process strictly follows the medical logic of tumor diagnosis, avoiding medically unreasonable recommendation results that may be generated by traditional recommendation algorithms. The adaptive learning characteristics of the Q-learning algorithm enable the system to dynamically adjust the recommendation strategy based on the historical recommendation performance of patients with different TNM stages, achieving truly personalized and accurate recommendations. The deep integration of temporal dynamic knowledge graphs and reinforcement learning algorithms not only ensures the dynamic adaptability of recommendation results in the time dimension, but also ensures a comprehensive balance between detection accuracy, cost-effectiveness, and time efficiency through multi-objective optimization. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a schematic diagram of an embodiment of the intelligent recommendation method for tumor reagents based on knowledge graphs in this application.

[0054] Figure 2 This is a schematic diagram of an embodiment of the intelligent recommendation process for tumor reagents based on knowledge graphs in this application.

[0055] Figure 3 This is a schematic diagram of an embodiment of the knowledge graph-based intelligent recommendation system for tumor reagents in this application.

[0056] Figure 4 This is a schematic block diagram of the intelligent recommendation device for tumor reagents based on knowledge graphs in an embodiment of the present invention. Detailed Implementation

[0057] This application provides a knowledge graph-based intelligent recommendation method, system, and storage medium for tumor reagents. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0058] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the intelligent recommendation method for tumor reagents based on knowledge graphs in this application includes:

[0059] Step S101: Add timestamp attributes to the tumor reagent node, biomarker node, and gene node, establish the correlation between the reagent detection accuracy attenuation coefficient and the biomarker expression fluctuation rate, and construct a time-series dynamic knowledge graph.

[0060] Step S102: When a reagent detection event occurs, update the memory state of the nodes in the temporal dynamic knowledge graph and generate a node temporal embedding vector containing dynamic features and time decay features.

[0061] Step S103: Obtain the patient's clinical characteristics, gene detection results and protein expression data, identify the patient's TNM staging information and output the patient state vector containing the scores of each staging;

[0062] Step S104: Calculate the time-weighted semantic similarity between the node temporal embedding vector and the patient state vector, and filter reagent nodes with a correlation degree higher than the threshold to obtain a candidate reagent set;

[0063] Step S105: Optimize the selection path of the candidate reagent set based on TNM staging constraints, and generate intelligent recommendation results including reagent identification, comprehensive score and recommendation basis.

[0064] It is understood that the executing entity of this application can be a knowledge graph-based intelligent recommendation system for tumor reagents, or it can be a terminal or a server; the specific implementation is not limited here. This application's embodiments use a server as an example for illustration.

[0065] Specifically, in the implementation of the intelligent recommendation system for tumor reagents, the construction of the time-series dynamic knowledge graph begins with adding time attributes. Reagent nodes are labeled with production dates and expiration dates, biomarker nodes record literature update times, and gene nodes store mutation detection dates. The detection accuracy decay coefficient is determined by fitting an exponential decay curve to accelerated aging experimental data, while the biomarker expression volatility is calculated from longitudinal data in the clinical testing database. The knowledge graph is stored using a time-series graph database, with time-series fields included in node and edge attributes, supporting graph query and update operations based on time ranges.

[0066] When a detection event triggers a map update, the event listener captures the reagent batch number, detection time, and performance indicators. The decay calculation engine updates the reagent node accuracy value based on the time interval and preset decay rate. The memory state module uses an LSTM network to fuse new data with historical states. The expression level of biomarker nodes is calculated using a fluctuation coefficient through time series analysis. This coefficient, together with the updated state of the reagent nodes, generates a temporal embedding vector containing static attributes, dynamic changes, and time decay features. The vector dimension design preserves the independent representation space of each feature.

[0067] The patient data processing workflow integrates electronic medical records, gene testing reports, and proteomics data. Clinical features are converted into feature vectors through medical terminology standardization, gene mutation information is quantified through variant annotation and functional prediction, and protein expression data undergoes batch correction and normalization. The TNM staging classifier receives multimodal feature input and simultaneously predicts tumor size, lymph node metastasis, and distant metastasis scores through a deep neural network, outputting a patient state vector containing continuous values, reflecting the quantitative characteristics of tumor progression.

[0068] Temporally weighted semantic similarity calculation projects the reagent time embedding vector and the patient state vector into a shared space. Cosine similarity measures the basic matching degree, and a time decay function calculates weight coefficients based on the reagent update time. The weighted similarity score is then used to filter candidate reagents through a preset threshold. The candidate set is optimized using a Q-learning algorithm constrained by TNM staging. The state space encodes the patient stage and candidate reagents, the action space defines the reagent selection strategy, and the reward function integrates three objectives: detection accuracy, cost-effectiveness, and time efficiency. The optimal recommendation path is obtained through iterative updates. The recommendation result generation module calculates a comprehensive score for reagent performance indicators, extracts supporting evidence from the knowledge graph, and outputs a structured report containing reagent identification, score details, and recommendation justification, ensuring clinical interpretability.

[0069] In one specific embodiment, the process of performing step S101 may specifically include the following steps:

[0070] We collected literature, reagent product instructions, and clinical databases in the field of oncology. We then cleaned and standardized the raw data to obtain a standardized basic dataset of oncology reagents.

[0071] Based on a standardized tumor reagent dataset, entity recognition and relation extraction are performed to obtain an entity set containing tumor reagent entities, biomarker entities, gene entities, and protein entities, as well as a set of relationships between entities.

[0072] Add a timestamp attribute and a decay coefficient attribute to each entity node in the entity set, and add a time weight attribute and an intensity change rate attribute to each relationship edge in the entity relationship set to obtain time-series labeled graph elements;

[0073] By storing the time-annotated spectral elements in a graph database, and establishing the decay function of reagent detection accuracy with time and the fluctuation function of biomarker expression level with time, a time-series dynamic knowledge graph is obtained.

[0074] Specifically, in constructing the time-series dynamic knowledge graph, oncology literature, reagent product instructions, and clinical databases serve as raw data inputs. A text cleaning process is used to remove non-standard terminology, redundant characters, and formatting differences. Text cleaning targets unstructured content in the literature, using regular expression matching to remove special symbols and standardizing reagent naming conventions, for example, standardizing "anti-PD-1 antibody" to "PD-1 detection reagent." Missing values ​​in the clinical database are filled using multiple imputation, and outliers are removed using boxplot detection. The processed data is converted into a unified structured format, and experimental data from the literature are mapped to reagent instructions using UUIDs, resulting in a standardized basic dataset of oncology reagents containing reagent names, detection indicators, and technical parameters.

[0075] Based on a standardized dataset of tumor reagents, the entity recognition process employs a bidirectional long short-term memory network model. The model input is a segmented text sequence, and the output layer uses CRF decoding to identify entities related to tumor reagents, biomarkers, genes, and proteins. During recognition, gene mutation data is converted to standard naming using the VEP annotation tool; for example, “7:140453136A>T” is annotated as “KRAS p.G12D”. The relation extraction process analyzes sentence dependency structure, extracting semantic relations such as “reagent-detection-marker” and “gene-encoding-protein”. Relation strength is quantified using a statistical p-value test. The processing results generate an entity set containing entities of tumor reagents, biomarkers, genes, and proteins, as well as a set of relationships between these entities.

[0076] Each entity node in the entity set is assigned a timestamp and a decay coefficient attribute. The timestamp records the data update time; for reagent nodes, the timestamp is derived from the production date corresponding to the product batch number, while for biomarker nodes, it is taken from the latest literature publication date. The decay coefficient attribute is calculated based on reagent stability experimental data, using the Arrhenius equation to predict the degradation rate at different temperatures. Relationship edges are assigned a time weight and an intensity change rate attribute. The time weight is calculated based on the validation frequency of the association, and the intensity change rate is determined using the sliding window standard deviation of historical data. The processed time-annotated spectral elements contain the temporal dynamics of the nodes and the intensity change characteristics of the edges.

[0077] Time-annotated spectral elements are stored in the Neo4j graph database, and a native graph storage engine is used to handle time-series queries of nodes and relationships. The reagent detection accuracy decay function over time is constructed as follows: Form, among which, For reagents in time The detection accuracy value at that time, The initial detection accuracy value of the reagent ( (Reference accuracy when =0) It is the natural constant (approximately 2.718). The attenuation coefficient / attenuation rate controls the rate at which precision decays. For time variables (time calculated from the start of reagent production or first use).

[0078] The parameters α and λ were obtained by fitting the experimental data of accelerated aging of the reagent, and the model was optimized using the least squares method.

[0079] The fluctuation function of the marker's level versus time is adopted form, For biomarkers in time The expression level value, The fluctuation amplitude coefficient determines the maximum deviation range that expresses horizontal fluctuations. It is a sine function used to simulate periodic fluctuations. Angular frequency controls the speed of the oscillation period. For time variables, The phase angle determines the initial phase position of the wave. Baseline level / mean value represents the baseline expression level of the biomarker.

[0080] Parameters were determined using Fourier analysis of clinical follow-up data. The function computation module updates node attribute values ​​in real time, ensuring that the knowledge graph dynamically reflects reagent performance degradation and biomarker expression fluctuations. The established time-series dynamic knowledge graph supports time-range-based path queries and near real-time update operations.

[0081] In one specific embodiment, the process of performing step S102 may specifically include the following steps:

[0082] The occurrence of reagent testing events in the time-series dynamic knowledge graph is monitored, and the event type, occurrence time, and performance data of the reagent testing events are identified and processed to obtain reagent testing event data.

[0083] Based on the reagent testing event data, the current detection accuracy of the relevant reagent nodes is calculated to reduce the accuracy. The updated detection accuracy value is calculated based on the time interval and the preset reduction rate to obtain the reagent node performance update data.

[0084] The performance update data of the reagent node is input into the memory state update module for node memory state refresh processing. At the same time, the expression level of the biomarker node is processed by fluctuation coefficient calculation to obtain the node update state containing performance decay information and expression fluctuation information.

[0085] Feature fusion processing is performed based on node update state, combining and encoding the node's inherent attributes, temporal change features, and time decay features to obtain the node time embedding vector.

[0086] Specifically, the time-series dynamic knowledge graph continuously monitors reagent testing events through a distributed event monitoring system. This system establishes a data interface with the hospital's laboratory information system, capturing operation records containing reagent batch numbers, testing times, and performance indicators in real time. The event type identification process parses the methodological descriptions in the testing reports, classifying and encoding detection technologies such as chemiluminescence and flow cytometry. Timestamp fields are synchronized to Coordinated Universal Time (UTC) using atomic clocks. Performance data, including detection limits and repeatability, is extracted from quality control reports. The raw event records undergo data transformation to generate structured reagent testing event data, where reagent identifiers are uniquely mapped to knowledge graph nodes, testing times are converted to Unix timestamp format, and performance parameters are processed for dimensional standardization and outlier filtering.

[0087] The reagent testing event data triggers the knowledge graph update process. The attenuation calculation process calculates the time interval between the node's last update and the timestamp information in the event, and queries the attenuation model parameters for the corresponding reagent category from a preset parameter table. The calculation process considers an environmental temperature correction factor, adjusting the attenuation rate using real-time temperature and humidity data obtained from the laboratory monitoring system. The updated detection accuracy value undergoes confidence interval calculation and validity verification, generating reagent node performance update data that includes the current accuracy value, measurement uncertainty, and expiration date.

[0088] The memory state update module employs a gated recurrent network architecture to process performance update data. The network input includes the node's current state vector, performance update data, and temporal context features. The update gating mechanism calculates the integration weights of new input information, and resetting the gating determines the retention ratio of historical memories, generating a new node memory state. Simultaneously, a biomarker expression fluctuation analysis is initiated, utilizing temporal detection data from a clinical database. Periodic fluctuation features are extracted using a sliding window Fourier transform, and an autoregressive model is employed to predict expression level trends. The processing results yield the node update state, where reagent nodes store precision decay curve parameters and memory cell states, while biomarker nodes record the mean, variance, and periodicity of expression levels.

[0089] The feature fusion stage inputs node intrinsic attribute vectors, memory state vectors, and time decay features into the feature combination network. The first layer of the network performs dimensionality transformation on static attributes, the second layer processes temporal memory features, and the third layer integrates decaying dynamic features. An attention mechanism calculates the relevance weights of each feature dimension, and a time-aware position encoder encodes absolute timestamps and relative time intervals into temporal feature vectors. The final generated node temporal embedding vector contains a 512-dimensional dense representation: the first 256 dimensions encode static semantic features, the middle 128 dimensions represent dynamic memory states, and the last 128 dimensions reflect time-sensitivity characteristics. This vector is stored through the extended attribute interface of the graph database, supporting similarity queries and path calculations based on time-sensitivity features.

[0090] In the data processing flow, the timestamps of detected events are synchronized with the global clock of the knowledge graph at the millisecond level, and performance metrics are uniformly converted to SI units. The memory state update network is trained using a contrastive loss function to ensure the separation of features with normal decay patterns and abnormal changes. The parameter initialization of the feature fusion network uses knowledge distillation technology to transfer and learn feature representation capabilities from a pre-trained medical language model. Layer normalization is introduced in the generation process of temporal embedding vectors to maintain the numerical stability of each feature dimension. The entire update process ensures data consistency through distributed transaction management, and changes in node states are recorded in the operation log for auditing and traceability.

[0091] In one specific embodiment, the process of executing step S103 may specifically include the following steps:

[0092] Collect basic patient information, medical history records and symptom descriptions, and perform standardized coding processing on the patient's clinical characteristics to obtain clinical feature vectors;

[0093] Mutation information and expression profile data are extracted from gene detection results, and gene mutation types and expression levels are quantitatively analyzed to obtain gene feature vectors.

[0094] The protein concentration and immunohistochemical results in the protein expression data were obtained, and the protein expression level was normalized to obtain the protein feature vector.

[0095] Clinical feature vectors, gene feature vectors, and protein feature vectors are input into the TNM staging classifier for staging identification, and tumor size score, lymph node involvement score, and distant metastasis score are calculated to obtain the patient status vector.

[0096] Specifically, basic patient information is extracted from structured fields in the hospital information system, including numerical data such as age, gender, and body surface area. These fields are scaled to the [0,1] interval using a min-max normalization method. Diagnostic codes in medical history records are mapped to the ICD-11 classification standard, surgical history and drug treatment history are converted into time-series event sequences, and symptom description text is used for named entity recognition using a clinical BERT model. After extracting key symptom terms, a 300-dimensional semantic embedding vector is generated. The standardized encoding of clinical features employs feature cross-validation technology, concatenating demographic data, medical history codes, and symptom vectors into a 512-dimensional clinical feature vector. Each dimension of the feature is sparsified using L1 regularization to eliminate the influence of redundant features.

[0097] Gene sequencing data processing involved obtaining VCF-formatted mutation data from the sequencing laboratory and performing functional annotation using the ANNOVAR tool to screen for non-synonymous mutations in exon regions. Mutation types were categorized into missense, nonsense, and splice site mutations, and converted into discrete feature vectors using One-Hot encoding. Expression profile data were extracted from RNA-seq quantification results, and after log2(x+1) transformation, batch effects were corrected using the ComBat method. The generation of gene feature vectors integrated mutation and expression features. Mutation load was quantified by calculating mutations per megabase. Driver gene status was labeled with clinical evidence levels based on the OncoKB database. Pathway activity was inferred from the expression profile using the ssGSEA algorithm. These features were then dimensionality-reduced using principal component analysis to obtain 256-dimensional gene feature vectors.

[0098] Protein expression data were acquired from the laboratory information management system. Protein intensity values ​​from mass spectrometry data were quantified using TMT reagent labeling and normalized using the median polish method. Immunohistochemical staining intensity scores were converted to continuous values, and the percentage of positive cells underwent logit transformation. The construction of protein feature vectors involved screening for biomarkers relevant to the current tumor type, such as ER, PR, and HER2 in breast cancer. These were then normalized using z-scores to generate 128-dimensional protein feature vectors. Lasso regression was used for feature selection, retaining protein indicators significantly correlated with clinical outcomes.

[0099] Clinical feature vectors, gene feature vectors, and protein feature vectors are input into the TNM staging classifier for processing. The classifier employs a multimodal fusion architecture. A feature crossover layer calculates the interaction features among clinical, gene, and protein features, while an attention mechanism layer dynamically adjusts the weights of each modality. Tumor size is determined by analyzing the maximum diameter data from CT images and the extent of invasion in pathology reports. Lymph node involvement is calculated based on the short axis diameter from ultrasound examination and the SUVmax value from PET-CT. Distant metastasis is integrated with circulating tumor cell detection results and whole-body bone scan reports. The classifier outputs a three-dimensional patient state vector containing T-staging, N-staging, and M-staging scores, ranging from 0 to 100, corresponding to the extent of the primary tumor, the degree of regional lymph node metastasis, and the status of distant metastasis, respectively.

[0100] During data processing, terminology standardization of clinical texts was performed using the UMLS meta-dictionary system, gene variant annotation referenced the latest version of the ClinVar database, and protein expression data underwent quality control according to the technical manual of the detection platform. The dimensional weights of feature vectors were optimized using elastic network regression, and the TNM classifier was trained using five-fold cross-validation. Evaluation metrics included accuracy and Cohen's kappa coefficient. The patient state vector update mechanism was designed as incremental learning; when new detection results were generated, only the affected feature dimensions were updated. This processing flow ensures that the patient state vector accurately reflects the anatomical extent and biological characteristics of the tumor, providing objective quantitative evidence for subsequent reagent recommendation algorithms.

[0101] In one specific embodiment, the process of executing step S104 may specifically include the following steps:

[0102] Extract the reagent feature component and time decay component from the node time embedding vector, and perform vector alignment processing on the clinical component and gene component in the patient state vector to obtain feature vector pairs;

[0103] The cosine similarity score is calculated based on the feature vector pairs. The time weighting coefficient is calculated based on the timestamp attribute of the reagent node. The cosine similarity score is then adjusted by time weighting to obtain the time-weighted semantic similarity score.

[0104] A threshold for the degree of association is set to a preset value. The time-weighted semantic similarity score is compared with the threshold, and reagent nodes with scores higher than the degree of association are selected to obtain a set of selected reagent nodes.

[0105] The set of reagent nodes is sorted in descending order according to the time-weighted semantic similarity score, and the top N reagent nodes are selected to obtain the candidate reagent set.

[0106] Specifically, the node temporal embedding vector is split into reagent feature components and time decay components through a feature separation layer. The reagent feature components contain static attribute encodings such as detection targets and reaction principles, while the time decay components store precision decay rate and environmental stability parameters. The clinical component in the patient state vector is projected to the same vector space as the reagent features through a fully connected layer, while the gene component undergoes dimensional alignment through pathway activity-based attention transformation. The vector alignment process employs a bilinear attention mechanism to calculate the interaction weights between each dimension of the patient features and the reagent features, generating aligned feature vector pairs through weighted combination. This process preserves the biological correlation between tumor molecular features and reagent detection principles.

[0107] The similarity measure for feature vector pairs employs an improved cosine similarity function, introducing a covariance matrix to correct feature distribution biases based on traditional dot product calculations. The derivation of the time weight coefficient is based on a reagent node timeliness model, inputting the time difference Δt between the current time and the node update time into a sigmoid decay function. Here, the parameter w is the time weight coefficient, k controls the decay rate, and τ represents the half-life threshold. The time-weighted adjustment process performs a Hadamard product operation on the semantic similarity matrix and the time weight tensor, resulting in higher weighted scores for recently validated reagent nodes while preserving basic matching information. Mathematically, this operation is equivalent to constructing a time-varying feature space, introducing dynamic decay constraints to static similarity.

[0108] The correlation threshold is determined using a dynamic programming algorithm, with a base value taken from a tumor type-staging table, and then individually adjusted based on the patient's treatment history and complications. The threshold comparison process employs a two-level screening architecture: first, potential candidates are quickly located using an approximate nearest neighbor search; then, a precise score is calculated for the streamlined node set. The generation of the screening reagent node set not only evaluates the matching degree of individual nodes but also detects the synergistic effect of related nodes in the knowledge graph using a random walk algorithm, ensuring that the screening results cover complementary detection schemes. This network-aware screening mechanism overcomes the technical limitation of traditional methods that neglect the synergistic effect between reagents.

[0109] The ranking process for the candidate reagent set employs a multi-objective optimization strategy. The primary ranking key is the time-weighted semantic similarity score, while the secondary ranking key considers the reagent's clinical validation level and cost-effectiveness ratio. The ranking algorithm introduces diversity constraints, ensuring that the top N reagents represent different technical approaches through maximum margin selection. The candidate set size N is dynamically configured based on TNM staging; N=10 is set for early-stage tumors to preserve exploration space, while N=5 is set for late-stage cases to focus on high-confidence schemes. The ranking results output a structured list containing reagent identifiers, matching scores, and timeliness markers, providing a pre-screened, high-quality candidate set for subsequent reinforcement learning.

[0110] In the data processing flow, the vector alignment projection matrix is ​​optimized using a contrastive loss function, and the time weight parameters k and τ are calibrated based on reagent stability experimental data. The dynamic adjustment of the correlation threshold uses clinically validated efficacy data as feedback signals, and the approximate search in the screening process is accelerated using locality-sensitive hashing. A diversity protection mechanism measures similarity through ontology distance of reagent action mechanisms, and the multi-objective weights of the ranking algorithm are defined by clinical experts.

[0111] In one specific embodiment, the process of executing step S105 may specifically include the following steps:

[0112] Define the state space to include a set of candidate reagents and a patient state vector, set the action space to include reagent selection action and path switching action, establish TNM staging constraints, and obtain the reinforcement learning environment configuration;

[0113] The candidate reagent set is input into the Q-learning algorithm for state value function initialization. The reward function weight parameters are set according to the TNM stage score in the patient's state vector to obtain the Q-value matrix.

[0114] Based on the Q-value matrix, an action selection strategy is implemented. The weighted sum of reagent testing accuracy reward, cost-effectiveness reward, and time-effect reward is calculated. The Q-value matrix is ​​then iteratively updated to obtain an optimized path sequence.

[0115] The optimal reagent selection scheme is extracted based on the optimized path sequence, and a comprehensive score is calculated for reagent detection accuracy, stability index and cost-effectiveness ratio. The result is then an intelligent recommendation that includes reagent identification, comprehensive score and selection criteria.

[0116] Specifically, the state space is constructed by mapping each reagent node in the candidate reagent set to a discrete state code, with encoding rules designed hierarchically based on reagent category and detection principle. The patient state vector is discretized into a finite set of states using the K-means clustering algorithm, with the number of cluster centers determined according to the number of TNM staging categories. The action space contains two basic operations: reagent selection actions correspond to specific reagent identifiers in the candidate set, and path switching actions define the adjustment strategy for the detection order. TNM staging constraints are implemented through a state transition matrix, with the values ​​of the matrix elements determined by tumor diagnosis and treatment guidelines, restricting the occurrence of reagent combinations that do not conform to staging characteristics. The reinforcement learning environment configuration also defines termination conditions, stopping the decision-making process when the recommended path length reaches a preset value or the cumulative reward exceeds a threshold.

[0117] The Q-learning algorithm's initialization process employs a dual-network architecture: the main network is responsible for Q-value prediction, while the target network provides a stable training objective. The input layer of the state-value function receives a 384-dimensional temporal embedding vector from reagent nodes and an 896-dimensional fusion vector from patient states. After dimensionality reduction through a three-layer fully connected network, the output is a Q-value estimate. The reward function weights are dynamically calculated using the TNM score from the patient state vector. Tumor size score primarily correlates with detection accuracy, lymph node score affects cost-effectiveness, and distant metastasis score determines temporal effectiveness. A temperature coefficient is introduced during weight allocation to adjust the exploration intensity; lower temperature values ​​are set for late-stage patients to enhance the utilization of known protocols. The Q-value matrix is ​​updated using a priority experience replay mechanism, adjusting the sample sampling probability based on temporal difference errors.

[0118] The action selection strategy initially employs Boltzmann exploration during training, transitioning to an ε-greedy strategy later. The accuracy reward is calculated based on CLIA-certified performance metrics, adjusted for confidence intervals using recent validation data. The cost-effectiveness reward integrates reagent prices, medical insurance reimbursement rates, and expected diagnostic value, standardized through quality-adjusted life years. The time-effect reward considers the match between the testing cycle and the rate of tumor progression, assigning higher sensitivity to urgent cases. The reward weighting uses a dynamic weighting mechanism, prioritizing accuracy in early decision steps and emphasizing cost considerations in later stages. Q-value updates utilize a double-Q learning algorithm to reduce overestimation bias, with the learning rate adaptively adjusted based on state access frequency.

[0119] The post-processing of the optimized pathway sequence filters out protocols that do not conform to clinical guidelines using predefined detection pathway templates. The extraction of the optimal reagent selection protocol employs Pareto front analysis, seeking non-dominated solutions within the three-dimensional objective space of accuracy, cost, and time. The comprehensive score is calculated using the analytic hierarchy process (AHP), transforming low-level indicators into global weights through a judgment matrix. Reagent detection accuracy indicators include the harmonic mean of sensitivity and specificity; stability indicators examine the inter-batch coefficient of variation; and cost-effectiveness ratio calculates the resource consumption per unit of diagnostic utility. The generation of intelligent recommendation results transforms structured decision data into textual descriptions conforming to clinical reporting standards, annotating the evidence level and source literature of key decision-making criteria. This processing workflow ensures that the recommendation results both meet the optimization objectives of reinforcement learning and satisfy the interpretability requirements of clinical practice.

[0120] In one specific embodiment, the process of inputting the candidate reagent set into the Q-learning algorithm for state value function initialization can specifically include the following steps:

[0121] Each reagent node identifier in the candidate reagent set is mapped to a state number, and the tumor size score, lymph node involvement score, and distant metastasis score in the patient state vector are normalized to obtain a state space coding table.

[0122] The Q-value function matrix is ​​initialized based on the state-space encoding table. The Q-value of each state-action pair in the matrix is ​​randomly initialized. The learning rate parameter and discount factor parameter are set to obtain the initial Q-value matrix.

[0123] Extract the TNM stage score values ​​from the patient state vector, calculate the detection accuracy weight, cost sensitivity weight, and time urgency weight based on the tumor severity, and normalize the weight parameters to obtain the set of reward function weight parameters.

[0124] By embedding the set of reward function weight parameters into the update formula of the initial Q-value matrix, the state transition probability function and the instantaneous reward function of the TNM stage constraint are constructed, and the Q-value matrix is ​​obtained.

[0125] Specifically, in the initialization phase of the Q-learning algorithm, node identifiers in the candidate reagent set are mapped to consecutive state numbers using a hash function. The hashing process preserves the hierarchical structure information of reagent categories, ensuring that reagents of the same type remain in adjacent positions in the state space. The TNM score component of the patient state vector undergoes min-max normalization, the tumor size score is converted into a discrete value in the [0,10] interval, the lymph node involvement score uses a logarithmic transformation to mitigate the influence of skewed distribution, and the distant metastasis score is graded and quantified according to the number of metastatic organs. The construction of the state space encoding table introduces a skip connection mechanism, which, in addition to the basic state number, records a combination key of reagent category code and patient staging characteristics, providing a fast index for subsequent state transitions.

[0126] The initialization of the Q-value function matrix employs a hierarchical strategy. The row dimension of the matrix corresponds to the total number of states in the state-space encoding table, while the column dimension covers all possible action-space combinations. The initial Q-value for each state-action pair is randomly sampled from a uniform distribution U(0, 0.1), with higher initial values ​​U(0.5, 0.6) assigned to state-action pairs corresponding to validated historical recommendation schemes in the knowledge graph. The learning rate parameter adopts a state-related decay strategy, with a base value of 0.1 that is dynamically adjusted based on the state access frequency. The discount factor parameter γ is determined to be 0.9 through eigenvalue analysis of the Bellman equation, achieving a balance between immediate rewards and long-term benefits. The storage structure of the initial Q-value matrix uses a sparse matrix compression format, storing only non-zero values ​​and their coordinate information to optimize memory usage efficiency.

[0127] The weighting of TNM staging scores employs a piecewise linear function transformation. Tumor size scores (T1-T4 interval) are mapped to detection accuracy weights of 0.2-0.8, lymph node scores (N0-N3 interval) to cost sensitivity weights of 0.1-0.7, and distant metastasis scores (M0-M1 interval) to time urgency weights of 0.3-0.9. The normalization of weight parameters incorporates a temperature coefficient τ to adjust the sharpness of the distribution; lower τ values ​​are set for advanced patients to enhance the weight differences of primary objectives. The update of the reward function weight parameter set uses a sliding window mechanism, retaining weight adjustment records from the most recent 10 decisions and smoothing parameter fluctuations through moving averages to avoid weight jumps caused by single assessments.

[0128] The state transition probability function integrates the structural constraints of the knowledge graph and the logical rules of clinical guidelines. The calculation of the transition probability P(s'|s,a) consists of three parts: the basic transition probability is derived from the statistical frequency of historical decision data; the knowledge graph connectivity adjustment term is calculated based on the path distance between nodes; and the clinical constraint term forces the transition probability of non-TNM staging requirements to be zero. The immediate reward function is defined using a multidimensional linear combination, multiplying the detection accuracy reward, cost-effectiveness reward, and time-effect reward by their respective weights and summing the results, then compressing them to the [-1,1] interval using the sigmoid function. The update formula for the Q-value matrix introduces the TNM staging-related advantage function A(s,a), adding a staging-related advantage term to the standard Q-learning update rule to accelerate the convergence of strategies that conform to the characteristics of the current staging. This initialization method that integrates domain knowledge solves the problem of cold start difficulties in traditional reinforcement learning in the medical field, shortens the training convergence time by integrating prior medical knowledge, and ensures that the recommended strategy conforms to the basic norms of tumor diagnosis and treatment.

[0129] The above describes the knowledge graph-based intelligent recommendation method for tumor reagents in the embodiments of this application. Below... Figure 2 The intelligent recommendation process for tumor reagents based on knowledge graphs in the embodiments of this application is described as follows:

[0130] Figure 2The knowledge graph-based intelligent recommendation process for tumor reagents, as shown, comprises five steps: construction of a time-series dynamic knowledge graph, generation of node time embedding vectors, generation of patient state vectors, screening of candidate reagent sets, and optimization of recommendation results. In the construction of the time-series dynamic knowledge graph, firstly, oncology literature, reagent product manuals, and clinical databases are collected. The raw data undergoes text cleaning and format standardization to obtain a standardized basic dataset of tumor reagents. Then, based on this dataset, entity recognition and relation extraction are performed to obtain a set containing entities of tumor reagents, biomarkers, genes, and proteins, as well as a set of relationships between entities. Subsequently, timestamps and decay coefficients are added to each entity node, and time weights and intensity change rate attributes are added to each relation edge to obtain time-annotated graph elements. Finally, these elements are stored in a graph database, establishing decay functions of reagent detection accuracy over time and fluctuation functions of biomarker expression levels over time. In the node temporal embedding vector generation process, reagent detection events are first monitored, and the event type, occurrence time, and performance data are identified to obtain reagent detection event data. Then, based on the event data, the detection accuracy decay of relevant reagent nodes is calculated to obtain reagent node performance update data. Next, the performance update data is input into the memory state update module to refresh the node memory state, and simultaneously, the fluctuation coefficient of biomarker node expression levels is calculated to obtain the node update state. Finally, feature fusion is performed on the node update state, combining and encoding the node's inherent attributes, temporal change features, and time decay features to generate the node temporal embedding vector. In the patient state vector generation process, patient clinical characteristics are first collected and standardized to obtain a clinical feature vector. Gene detection results are extracted and quantitatively analyzed to obtain a gene feature vector, and protein expression data is obtained and normalized to obtain a protein feature vector. Then, the three vectors are input into the TNM staging classifier to calculate tumor size, lymph node involvement, and distant metastasis scores to obtain the patient state vector. In the candidate reagent set screening process, reagent features and time decay components are first extracted from the node time embedding vectors. The clinical and genetic components of the patient state vectors are then aligned to obtain feature vector pairs. Next, cosine similarity is calculated and combined with a time weight coefficient to obtain a time-weighted semantic similarity score. Then, reagent nodes with scores higher than a threshold are selected, and the top N nodes are chosen in descending order of score to obtain the candidate reagent set. In the recommendation result optimization and generation process, the state space, action space, and TNM staging constraints are first defined, and a reinforcement learning environment is configured. Then, the candidate reagent set is input into the Q-learning algorithm to initialize the state value function. The reward function weight parameters are set according to the TNM staging score to obtain the Q-value matrix. Next, an action selection strategy is executed based on the Q-value matrix, iteratively updating the Q-value matrix to obtain an optimized path sequence. Finally, the optimal solution is extracted, a comprehensive score is calculated, and an intelligent recommendation result containing reagent identification, comprehensive score, and recommendation criteria is generated.

[0131] The above describes the knowledge graph-based intelligent recommendation process for tumor reagents in the embodiments of this application. The following describes the knowledge graph-based intelligent recommendation system 200 for tumor reagents in the embodiments of this application. Please refer to [link to relevant documentation]. Figure 3 One embodiment of the knowledge graph-based intelligent recommendation system for tumor reagents 200 in this application includes:

[0132] Module 201 is established to add timestamp attributes to tumor reagent nodes, biomarker nodes, and gene nodes, establish the correlation between reagent detection accuracy attenuation coefficient and biomarker expression volatility, and construct a time-series dynamic knowledge graph.

[0133] The generation module 202 is used to update the memory state of nodes in the time-series dynamic knowledge graph when a reagent detection event occurs, and generate node time embedding vectors containing dynamic features and time decay features.

[0134] The identification module 203 is used to acquire the patient's clinical characteristics, gene detection results and protein expression data, and after identifying the patient's TNM staging information, outputs a patient state vector containing the scores of each staging.

[0135] The filtering module 204 is used to calculate the time-weighted semantic similarity between the node time embedding vector and the patient state vector, and filter reagent nodes with a correlation degree higher than a threshold to obtain a candidate reagent set.

[0136] The optimization module 205 is used to optimize the selection path of the candidate reagent set based on TNM staging constraints, and generate intelligent recommendation results including reagent identification, comprehensive score and recommendation basis.

[0137] above Figure 3 The knowledge graph-based intelligent recommendation system for tumor reagents in this embodiment of the invention is described in detail from the perspective of modular functional entities. The knowledge graph-based intelligent recommendation device for tumor reagents in this embodiment of the invention is described in detail from the perspective of hardware processing.

[0138] Reference Figure 4 This invention also provides a knowledge graph-based intelligent recommendation device 300 for tumor reagents. This knowledge graph-based intelligent recommendation device for tumor reagents can be a server, and its internal structure can be as follows: Figure 4As shown. This knowledge graph-based intelligent recommendation device for tumor reagents includes a processor 302, a memory 303, a display screen 304, an input device 305, a network interface 306, and a database 307 connected via a system bus 301. The processor 302, designed as a computer, provides computational and control capabilities. The memory 303 of this knowledge graph-based intelligent recommendation device includes a non-volatile storage medium 3031 and internal memory 3032. The non-volatile storage medium 3031 stores the operating system and computer programs. The internal memory 3032 provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database 307 of this knowledge graph-based intelligent recommendation device stores the data corresponding to this embodiment. The network interface 306 of this knowledge graph-based intelligent recommendation device is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the above-described method.

[0139] Those skilled in the art will understand that Figure 4 The structures shown are merely block diagrams of some structures related to the present invention and do not constitute a limitation on the knowledge graph-based intelligent recommendation device for tumor reagents applied thereto.

[0140] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the knowledge graph-based intelligent recommendation method for tumor reagents.

[0141] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0142] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a knowledge graph-based intelligent recommendation device for tumor reagents (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0143] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A knowledge graph-based intelligent recommendation method for tumor reagents, characterized in that, The method includes: Step S101: Add timestamp attributes to the tumor reagent node, biomarker node, and gene node, establish the correlation between the reagent detection accuracy attenuation coefficient and the biomarker expression fluctuation rate, and construct a time-series dynamic knowledge graph. Step S102: Monitor the occurrence of reagent detection events in the time-series dynamic knowledge graph, identify and process the event type, occurrence time, and performance data of the reagent detection events to obtain reagent detection event data; perform attenuation calculation processing on the current detection accuracy of relevant reagent nodes based on the reagent detection event data, calculate the updated detection accuracy value based on the time interval and preset attenuation rate, and obtain reagent node performance update data; input the reagent node performance update data into the memory state update module for node memory state refresh processing, and simultaneously calculate the fluctuation coefficient processing of the expression level of biomarker nodes to obtain node update state containing performance attenuation information and expression fluctuation information; perform feature fusion processing based on the node update state, and combine and encode the node's inherent attributes, temporal change features, and time attenuation features to obtain the node temporal embedding vector; Step S103: Obtain the patient's clinical characteristics, gene detection results and protein expression data, identify the patient's TNM staging information and output the patient state vector containing the scores of each staging; Step S104: Extract the reagent feature component and time decay component from the node time embedding vector; perform vector alignment processing on the clinical component and gene component in the patient state vector to obtain feature vector pairs; calculate the cosine similarity value based on the feature vector pairs; calculate the time weight coefficient according to the timestamp attribute of the reagent node; perform time weighting adjustment processing on the cosine similarity value to obtain a time-weighted semantic similarity score; set the association degree threshold to a preset value; perform threshold comparison processing on the time-weighted semantic similarity score; filter reagent nodes with scores higher than the association degree threshold to obtain a set of filtered reagent nodes; sort the set of filtered reagent nodes in descending order according to the time-weighted semantic similarity score; select the top N reagent nodes to obtain a set of candidate reagents. Step S105: Define a state space containing the candidate reagent set and the patient state vector; set an action space containing reagent selection actions and path switching actions; establish TNM staging constraints to obtain the reinforcement learning environment configuration; input the candidate reagent set into the Q-learning algorithm for state value function initialization; set reward function weight parameters according to the TNM staging score in the patient state vector to obtain the Q-value matrix; execute the action selection strategy based on the Q-value matrix; calculate the weighted sum of reagent detection accuracy reward, cost-effectiveness reward, and time effect reward; iteratively update the Q-value matrix to obtain the optimized path sequence; extract the optimal reagent selection scheme according to the optimized path sequence; calculate the comprehensive score of reagent detection accuracy, stability index, and cost-effectiveness ratio; and generate the intelligent recommendation result containing reagent identification, the comprehensive score, and selection criteria.

2. The intelligent recommendation method for tumor reagents based on knowledge graphs according to claim 1, characterized in that, Step S101 includes: We collected literature, reagent product instructions, and clinical databases in the field of oncology. We then cleaned and standardized the raw data to obtain a standardized basic dataset of oncology reagents. Based on the standardized tumor reagent dataset, entity recognition and relation extraction are performed to obtain an entity set containing tumor reagent entities, biomarker entities, gene entities, and protein entities, as well as a set of relationships between entities. Add a timestamp attribute and a decay coefficient attribute to each entity node in the entity set, and add a time weight attribute and an intensity change rate attribute to each relationship edge in the entity relationship set to obtain time-annotated graph elements; The time-annotated spectral elements are stored in a graph database, and a decay function of reagent detection accuracy with time and a fluctuation function of biomarker expression level with time are established to obtain the time-series dynamic knowledge graph.

3. The intelligent recommendation method for tumor reagents based on knowledge graphs according to claim 1, characterized in that, Step S103 includes: Collect basic patient information, medical history records, and symptom descriptions, and perform standardized coding processing on the patient's clinical characteristics to obtain a clinical feature vector; Mutation information and expression profile data are extracted from the gene detection results, and the gene mutation types and expression levels are quantitatively analyzed to obtain gene feature vectors. The protein concentration and immunohistochemical results in the protein expression data are obtained, and the protein expression level is normalized to obtain a protein feature vector. The clinical feature vector, gene feature vector and protein feature vector are input into the TNM staging classifier for staging identification, and the tumor size score, lymph node involvement score and distant metastasis score are calculated to obtain the patient status vector.

4. The intelligent recommendation method for tumor reagents based on knowledge graphs according to claim 1, characterized in that, The process involves inputting the candidate reagent set into a Q-learning algorithm for state value function initialization, setting reward function weight parameters based on the TNM staging score in the patient's state vector, and obtaining a Q-value matrix, including: Each reagent node identifier in the candidate reagent set is mapped to a state number, and the tumor size score, lymph node involvement score, and distant metastasis score in the patient state vector are normalized to obtain a state space coding table. The Q-value function matrix is ​​initialized based on the state space encoding table. The Q-value of each state-action pair in the matrix is ​​randomly initialized. The learning rate parameter and discount factor parameter are set to obtain the initial Q-value matrix. Extract the TNM stage score values ​​from the patient state vector, calculate the detection accuracy weight, cost sensitivity weight, and time urgency weight according to the tumor severity, and perform normalization constraint processing on the weight parameters to obtain the set of reward function weight parameters; The reward function weight parameter set is embedded into the update formula of the initial Q-value matrix to construct the state transition probability function and the instantaneous reward function of the TNM phased constraint, thereby obtaining the Q-value matrix.

5. A knowledge graph-based intelligent recommendation system for tumor reagents, characterized in that, For implementing the knowledge graph-based intelligent recommendation method for tumor reagents as described in any one of claims 1-4, the knowledge graph-based intelligent recommendation system for tumor reagents comprises: A module was established to add timestamp attributes to tumor reagent nodes, biomarker nodes, and gene nodes, establish the correlation between reagent detection accuracy attenuation coefficient and biomarker expression volatility, and construct a time-series dynamic knowledge graph. The generation module is used to update the memory state of nodes in the time-series dynamic knowledge graph when a reagent detection event occurs, and generate node time embedding vectors containing dynamic features and time decay features. The identification module is used to acquire patients’ clinical characteristics, gene testing results and protein expression data, and after identifying the patient’s TNM staging information, it outputs a patient state vector containing the scores of each staging. The filtering module is used to calculate the time-weighted semantic similarity between the node time embedding vector and the patient state vector, and filter reagent nodes with a correlation degree higher than a threshold to obtain a candidate reagent set. The optimization module is used to optimize the selection path of the candidate reagent set based on TNM staging constraints, and generate intelligent recommendation results including reagent identification, comprehensive score and recommendation basis.

6. A knowledge graph-based intelligent recommendation device for tumor reagents, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the knowledge graph-based intelligent recommendation method for tumor reagents as described in any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it causes the processor to execute the knowledge graph-based intelligent recommendation method for tumor reagents as described in any one of claims 1 to 4.

Images